When Anthropic launched the AI Mythos model in April, the company issued a notable warning that the new system is so powerful that it can detect thousands of serious vulnerabilities in software before being widely publicized.
Currently, Mozilla (the developer of the Firefox browser) has revealed more clearly how Mythos is completely changing their cybersecurity process.
In a newly published post, Mozilla said Mythos has discovered a large number of serious errors in Firefox, including vulnerabilities that have been hidden in the source code for more than 10 years. This is considered a very big step forward compared to previous AI security tools that often created too many false alarms or low-quality reports.
According to Mozilla researchers, in just a few months, the efficiency of AI has improved significantly. Not only have models become more accurate, but the technology of exploiting AI to scan for errors has also been strongly upgraded.
Systems today can self-assess their own results and eliminate unreliable findings.
The impact of this technology is clearly shown through the number of patches. In April 2026, Firefox released up to 423 security patches, while in the same period last year there were only 31 patches.
Mozilla also announced details of 12 serious bugs, including two rare vulnerabilities in the "sandbox" mechanism and a bug that has existed for 15 years related to how the browser processes HTML elements.
Brian Grinstead, senior engineer at Mozilla, said that current AI tools have progressed far beyond expectations. According to him, Mozilla has seen clear changes through internal scanning, external reports as well as the general trend of the entire security industry.
In particular, the ability to detect errors in sandbox environments is considered the most impressive achievement of Mythos.
Sandbox is an important security layer that helps isolate malware and prevent widespread attacks in the browser.
Detecting vulnerabilities in this area requires a very complex process, including many steps such as creating a test patch, deploying new code, and then continuing to check exploitation capabilities.
Mozilla is currently offering a reward of up to $20,000 USD for experts who find Firefox sandbox errors. However, Grinstead admits Mythos is discovering more of this type of error than previous human security researchers.
Although AI supports strongly in finding errors, Mozilla has not yet fully assigned the correction of errors to artificial intelligence.
According to the Firefox development team, the patches created by AI currently only serve as references because human engineers still need to directly write and check the source code before deployment.
The development of Mythos has also sparked major debate in the cybersecurity industry. Some experts worry that this technology could be exploited by hackers to search for vulnerabilities faster.
However, Anthropic believes that in the long term AI will help the defense side gain an advantage thanks to its ability to detect and correct errors on a large scale.
Anthropic CEO - Dario Amodei, said that if used properly, AI can help create a safer software environment than before.
Meanwhile, Mozilla believes that the race between AI attack and defense has not yet had a final answer, but currently this technology is helping the balance lean slightly towards system protection.
